ASPS v2: Mission Policy Specification

Version: 2.0-draft Status: Draft Date: 2026-04-09 Extends: ASPS v1: Agent Spending Policy

Abstract

ASPS v2 extends ASPS v1 to support coordinated spending across multiple AI agents working toward a common goal. While v1 defines rules for a single agent, v2 defines rules for a mission — a task with shared resources, phases, dependencies, dynamic budget allocation, and risk scoring.

ASPS v2 is a layer above v1, not a replacement. Every agent inside a mission still uses an ASPS v1 policy for individual spending rules.

1. Motivation

AI agents increasingly work in teams. A travel planning task may involve 5 agents (flights, hotels, activities, research, logistics). A marketing campaign may involve 3 agents (content, ads, analytics). An incident response may involve 4 agents (monitoring, diagnosis, remediation, postmortem).

These teams share budgets, have dependencies between tasks, progress through phases, and need coordinated spending control. No existing specification addresses this.

2. Core Concepts

2.1 Mission

A mission is a goal-oriented task with a shared resource pool and a set of participating agents.

{
  "version": "2.0",
  "name": "Barcelona trip, May 15-22",
  "budget": 5000,
  "currency": "USD",
  "deadline": "2026-05-10T00:00:00Z",
  "agents": { ... },
  "phases": [ ... ],
  "constraints": [ ... ],
  "on_failure": "pause"
}

Field	Type	Required	Description
`version`	`string`	No	Spec version
`name`	`string`	Yes	Human-readable mission name
`budget`	`number`	Yes	Total mission budget
`currency`	`string`	Yes	ISO 4217 currency code
`deadline`	`datetime`	No	Mission deadline (ISO 8601)
`agents`	`object`	Yes	Agent definitions (keyed by role)
`phases`	`Phase[]`	Yes	Ordered list of mission phases
`constraints`	`Constraint[]`	No	Cross-agent and cross-phase constraints
`on_failure`	`enum`	No	What happens when a phase fails: `pause`, `rollback`, `abort`. Default: `pause`

2.2 Agent Role

Each agent in a mission has a role and an optional individual ASPS policy.

{
  "agents": {
    "flights": {
      "description": "Searches and books flights",
      "policy": {
        "allowed_categories": ["flights", "transport"],
        "per_request_limit": 1500
      }
    },
    "hotel": {
      "description": "Searches and books accommodation",
      "policy": {
        "allowed_categories": ["accommodation"],
        "per_request_limit": 2000
      }
    },
    "entertainment": {
      "description": "Plans and books activities",
      "policy": {
        "allowed_categories": ["entertainment", "restaurants", "transport"]
      }
    },
    "research": {
      "description": "Weather, reviews, safety analysis",
      "can_spend": false
    }
  }
}

Field	Type	Required	Description
`description`	`string`	No	What this agent does
`policy`	`Policy`	No	ASPS policy for this agent within the mission
`can_spend`	`boolean`	No	If `false`, agent participates but cannot make spending requests. Default: `true`

The agent-level ASPS policy (if set) is applied in addition to phase-level and mission-level rules. All must pass.

2.3 Phase

A phase is a stage of the mission with its own budget allocation, active agents, and entry/exit conditions.

{
  "phases": [
    {
      "name": "research",
      "agents": ["research", "flights", "hotel"],
      "allocation": { "type": "fixed", "amount": 0 },
      "exit_condition": { "type": "manual" }
    },
    {
      "name": "booking",
      "agents": ["flights", "hotel"],
      "allocation": { "type": "share", "percent": 70 },
      "exit_condition": {
        "type": "all_confirmed",
        "agents": ["flights", "hotel"]
      }
    },
    {
      "name": "activities",
      "agents": ["entertainment"],
      "allocation": { "type": "remaining" },
      "exit_condition": { "type": "manual" }
    }
  ]
}

Field	Type	Required	Description
`name`	`string`	Yes	Phase identifier
`agents`	`string[]`	Yes	Agent roles active in this phase
`allocation`	`Allocation`	Yes	How budget is allocated to this phase
`exit_condition`	`ExitCondition`	No	When this phase is considered complete
`on_failure`	`enum`	No	Override mission-level `on_failure` for this phase
`timeout`	`string`	No	Max duration for this phase (e.g. `"24h"`, `"7d"`)

Phases are ordered. A phase starts when the previous phase's exit_condition is met. The first phase starts immediately.

Multiple phases MAY be active simultaneously if explicitly configured (see parallel in Section 5).

3. Budget Allocation

3.1 Allocation Types

Type	Description	Example
`fixed`	Fixed dollar amount	`{ "type": "fixed", "amount": 500 }`
`share`	Percentage of total mission budget	`{ "type": "share", "percent": 70 }`
`remaining`	Whatever is left after previous phases	`{ "type": "remaining" }`
`per_agent`	Fixed amount per agent in the phase	`{ "type": "per_agent", "amount": 100 }`
`competitive`	Initial amount per agent, winner gets the rest	`{ "type": "competitive", "seed": 100, "prize": 1000, "metric": "lowest_price" }`

3.2 Dynamic Reallocation

Within a phase, unspent budget from one agent is available to others. This is the default behavior.

{
  "allocation": {
    "type": "share",
    "percent": 70,
    "reallocation": "dynamic"
  }
}

Reallocation	Description
`dynamic`	Unspent budget is available to other agents in the phase (default)
`partitioned`	Each agent gets a fixed slice, no reallocation

With dynamic reallocation: if flights agent spends $800 of a $3500 phase budget, hotel agent can spend up to $2700.

With partitioned: flights gets $1750, hotel gets $1750, regardless of what each spends.

3.3 Budget Cascade

Budget flows top-down:

Mission budget ($5000)
  → Phase allocation (research: $0, booking: $3500, activities: remaining)
    → Agent policy (per_request_limit, categories, etc.)
      → ASPS checks (9 standard checks)

A request must pass ALL levels. Mission budget exhausted → reject, even if phase and agent budgets allow it.

4. Transitions and Exit Conditions

4.1 Exit Condition Types

Type	Description	Example
`manual`	Human explicitly advances to next phase	`{ "type": "manual" }`
`all_confirmed`	All specified agents have confirmed purchases	`{ "type": "all_confirmed", "agents": ["flights", "hotel"] }`
`any_confirmed`	At least one agent has confirmed	`{ "type": "any_confirmed", "agents": ["agent_a", "agent_b"] }`
`budget_depleted`	Phase budget is fully spent	`{ "type": "budget_depleted" }`
`timeout`	Phase duration exceeded	Configured via `timeout` field on Phase
`condition`	Custom expression	`{ "type": "condition", "expr": "flights.spent > 0 AND hotel.spent > 0" }`

4.2 Phase Transitions

When a phase's exit condition is met:

Current phase is marked completed
Unspent phase budget returns to the mission pool (available for remaining allocations)
Next phase starts, its agents become active
Agents not in the new phase are paused (cannot make requests)

4.3 Failure Handling

When a phase fails (timeout, constraint violation, explicit failure signal):

on_failure	Behavior
`pause`	Mission pauses. Human decides next step. (Default)
`rollback`	Engine emits rollback signals for confirmed purchases in the failed phase. Previous phase reactivates.
`abort`	Mission terminates. All pending requests are cancelled. Holds are released.

Rollback is a signal, not an automatic refund. The engine tells agents "this purchase should be reversed." The actual reversal is the agent's responsibility.

5. Constraints

Constraints express relationships between agents and phases.

{
  "constraints": [
    {
      "type": "dependency",
      "agent": "hotel",
      "requires": "flights",
      "condition": "confirmed"
    },
    {
      "type": "combined_limit",
      "agents": ["flights", "hotel"],
      "max_share": 0.7
    },
    {
      "type": "conditional_limit",
      "if": "hotel.last_amount > 200",
      "then": { "agent": "entertainment", "daily_limit": 50 }
    }
  ]
}

5.1 Constraint Types

Type	Description
`dependency`	Agent B cannot spend until Agent A meets a condition
`combined_limit`	Sum of spending across listed agents cannot exceed a limit
`conditional_limit`	If a condition is true, apply additional limits to an agent
`exclusion`	If Agent A has spent, Agent B is blocked (mutually exclusive)
`priority_order`	Agents spend in priority order; lower-priority agents wait

5.2 Dependency Constraint

{
  "type": "dependency",
  "agent": "hotel",
  "requires": "flights",
  "condition": "confirmed"
}

Condition	Meaning
`confirmed`	Required agent has at least one confirmed purchase
`approved`	Required agent has at least one approved request
`phase_complete`	Required agent's phase is complete
`spent_above`	Required agent has spent above a threshold: `"condition": "spent_above:500"`

5.3 Combined Limit Constraint

{
  "type": "combined_limit",
  "agents": ["flights", "hotel"],
  "max_share": 0.7
}

Sum of all spending by listed agents cannot exceed max_share (as fraction of mission budget) or max_amount (absolute value).

5.4 Conditional Limit Constraint

{
  "type": "conditional_limit",
  "if": "hotel.last_amount > 200",
  "then": { "agent": "entertainment", "daily_limit": 50 }
}

If the condition evaluates to true, the then policy is merged with the agent's existing policy (stricter values win).

6. Escalation Model

For scenarios where agents gain more authority over time (customer service, incident response).

{
  "phases": [
    {
      "name": "L1",
      "agents": ["support"],
      "allocation": { "type": "fixed", "amount": 50 },
      "exit_condition": { "type": "manual" }
    },
    {
      "name": "L2",
      "agents": ["support"],
      "allocation": { "type": "fixed", "amount": 500 },
      "exit_condition": { "type": "manual" }
    },
    {
      "name": "L3",
      "agents": ["support"],
      "allocation": { "type": "fixed", "amount": 5000 },
      "exit_condition": { "type": "manual" }
    }
  ]
}

In escalation, the same agent participates in multiple phases with increasing budgets. Each phase is a new level of authority. Transition is typically manual (a human or system escalates).

Note: in escalation, phase budgets are not cumulative — each level has its own allocation. Unspent L1 budget does not carry to L2.

7. Competitive Allocation

For scenarios where agents compete and the winner gets the budget.

{
  "phases": [
    {
      "name": "search",
      "agents": ["supplier_a", "supplier_b", "supplier_c"],
      "allocation": {
        "type": "competitive",
        "seed": 50,
        "prize": 2000,
        "metric": "lowest_price"
      },
      "timeout": "24h"
    }
  ]
}

Field	Type	Description
`seed`	`number`	Initial budget per agent for research/exploration
`prize`	`number`	Budget awarded to the winning agent
`metric`	`string`	How to pick the winner: `lowest_price`, `highest_quality`, `manual`

When the phase ends (timeout or manual trigger), the engine evaluates results by metric and awards prize to the winning agent. Losing agents are paused.

8. Mission Lifecycle

CREATED → ACTIVE → COMPLETED
              ↓
           PAUSED → ACTIVE (resume)
              ↓
           ABORTED

8.1 States

State	Description
`created`	Mission defined but not started
`active`	At least one phase is active, agents can make requests
`paused`	Mission paused (failure, manual intervention). No requests accepted
`completed`	All phases completed successfully
`aborted`	Mission terminated. All holds released

8.2 Request Evaluation Within a Mission

When an agent submits a spending request:

Mission state — mission must be active
Phase check — agent must be in the currently active phase
Agent role — agent must have can_spend: true
Phase budget — request must fit within phase allocation (considering dynamic reallocation)
Mission budget — request must fit within total mission budget
Constraints — all applicable constraints must pass (dependencies, combined limits, conditional limits)
Agent ASPS policy — standard 9 ASPS checks against agent's individual policy
Risk check — if risk policy is configured (see ASPS risk scoring)

All checks must pass. Failure at any level → reject with detailed report.

9. Risk Scoring

ASPS v2 adds an optional risk scoring check that detects anomalous spending patterns. Unlike deterministic checks (limits, categories), risk scoring evaluates whether a request is unusual for this agent, even if it formally passes all rules.

9.1 RiskPolicy Object

Added to the agent's ASPS v1 policy or to the mission-level configuration.

{
  "risk": {
    "risk_threshold": 0.7,
    "on_high_risk": "pending",
    "baseline_window": "30d"
  }
}

Field	Type	Required	Description
`risk_threshold`	`number`	Yes	Score threshold (0.0–1.0). Requests scoring above this are flagged
`on_high_risk`	`enum`	No	Action when threshold exceeded: `pending` or `reject`. Default: `pending`
`baseline_window`	`string`	No	Period for building baseline behavior (e.g. `"30d"`, `"7d"`). Default: `"30d"`

9.2 Risk Check

Evaluated as the last check (after ASPS v1 checks and mission-level checks). The specification defines the interface, not the algorithm — implementations choose how to compute the score.

Recommended risk factors (implementations SHOULD consider):

Factor ID	Description
`unusual_amount`	Amount significantly deviates from agent's typical spending in this category
`unusual_category`	Agent has not previously used this category
`unusual_time`	Request at an atypical time for this agent
`unusual_frequency`	Atypical request frequency (e.g. 3 requests/minute vs usual 2/hour)
`description_mismatch`	Description does not semantically match the category

Implementations MAY add custom factors.

9.3 Risk Check Result

{
  "rule": "risk",
  "result": "fail",
  "detail": "Risk score 0.82 exceeds threshold 0.7",
  "risk_score": 0.82,
  "risk_factors": ["unusual_amount", "unusual_time"]
}

The risk_score and risk_factors fields extend the standard Check Result object (ASPS v1 Section 8).

9.4 Implementation Approaches

The specification does not mandate an algorithm. Possible approaches (from simple to complex):

Statistical (z-score) — deviation from mean per factor
Rule-based — weighted heuristics
ML model — trained on agent's transaction history

10. Relationship to ASPS v1

ASPS v2 is a layer above v1, not a replacement.

ASPS v2 (missions, coordination, risk scoring)
  └── ASPS v1 (per-agent spending rules)
        └── 9 standard checks

An agent inside a mission still has an ASPS v1 policy
ASPS v1 checks are evaluated as part of step 7 (see Section 8.2)
An agent can operate independently (with just ASPS v1) or as part of a mission
Account-level Budget Rules (ASPS v1 Section 6) apply across all missions
Risk scoring (Section 9) can be used with or without missions

{
  "version": "2.0",
  "name": "Barcelona trip, May 15-22, 2 people",
  "budget": 5000,
  "currency": "USD",
  "deadline": "2026-05-10T00:00:00Z",

  "agents": {
    "research": {
      "description": "Weather, reviews, safety, geopolitics",
      "can_spend": false
    },
    "flights": {
      "description": "Search and book flights",
      "policy": {
        "allowed_categories": ["flights", "transport"],
        "per_request_limit": 1500
      }
    },
    "hotel": {
      "description": "Search and book accommodation",
      "policy": {
        "allowed_categories": ["accommodation"],
        "per_request_limit": 2000
      }
    },
    "entertainment": {
      "description": "Plan and book activities, restaurants",
      "policy": {
        "allowed_categories": ["entertainment", "restaurants", "transport"],
        "per_request_limit": 200
      }
    }
  },

  "phases": [
    {
      "name": "research",
      "agents": ["research", "flights", "hotel"],
      "allocation": { "type": "fixed", "amount": 0 },
      "exit_condition": { "type": "manual" }
    },
    {
      "name": "booking",
      "agents": ["flights", "hotel"],
      "allocation": { "type": "share", "percent": 70, "reallocation": "dynamic" },
      "exit_condition": {
        "type": "all_confirmed",
        "agents": ["flights", "hotel"]
      }
    },
    {
      "name": "activities",
      "agents": ["entertainment"],
      "allocation": { "type": "remaining" },
      "exit_condition": { "type": "manual" }
    }
  ],

  "constraints": [
    {
      "type": "dependency",
      "agent": "hotel",
      "requires": "flights",
      "condition": "approved"
    },
    {
      "type": "combined_limit",
      "agents": ["flights", "hotel"],
      "max_share": 0.75
    }
  ],

  "on_failure": "pause"
}

This mission:

Total budget $5,000 for the trip
Research phase: agents explore options, no spending
Booking phase: 70% of budget ($3,500), dynamic reallocation between flights and hotel
Hotel cannot book until flights are at least approved
Flights + hotel combined cannot exceed 75% of total ($3,750)
Activities phase: gets whatever remains after booking
If anything fails, mission pauses for human decision

Appendix B: Full Example — Incident Response (Escalation)

{
  "version": "2.0",
  "name": "Production outage #4521",
  "budget": 10000,
  "currency": "USD",

  "agents": {
    "monitor": {
      "description": "Detects and reports issues",
      "can_spend": false
    },
    "diagnose": {
      "description": "Investigates root cause",
      "policy": {
        "allowed_categories": ["compute", "diagnostics"]
      }
    },
    "remediate": {
      "description": "Fixes the issue",
      "policy": {
        "allowed_categories": ["compute", "networking", "storage"]
      }
    }
  },

  "phases": [
    {
      "name": "detection",
      "agents": ["monitor"],
      "allocation": { "type": "fixed", "amount": 0 },
      "exit_condition": { "type": "manual" }
    },
    {
      "name": "diagnosis",
      "agents": ["diagnose"],
      "allocation": { "type": "fixed", "amount": 500 },
      "timeout": "2h",
      "exit_condition": { "type": "manual" }
    },
    {
      "name": "remediation",
      "agents": ["remediate"],
      "allocation": { "type": "fixed", "amount": 5000 },
      "timeout": "4h",
      "exit_condition": { "type": "manual" }
    },
    {
      "name": "escalation",
      "agents": ["remediate"],
      "allocation": { "type": "remaining" },
      "exit_condition": { "type": "manual" }
    }
  ],

  "on_failure": "pause"
}

Appendix C: Full Example — Competitive Search

{
  "version": "2.0",
  "name": "Find best laptop deal",
  "budget": 2500,
  "currency": "USD",

  "agents": {
    "amazon": { "description": "Searches Amazon" },
    "ebay": { "description": "Searches eBay" },
    "bestbuy": { "description": "Searches Best Buy" }
  },

  "phases": [
    {
      "name": "search",
      "agents": ["amazon", "ebay", "bestbuy"],
      "allocation": {
        "type": "competitive",
        "seed": 0,
        "prize": 2500,
        "metric": "lowest_price"
      },
      "timeout": "12h",
      "exit_condition": { "type": "timeout" }
    }
  ]
}